A Deep Learning-Based Model for Nowcasting of Convective Initiation Using Infrared Observations

As severe convective weather exerts growing influence on public safety, enhancing forecast accuracy has become critically important. However, the predictive capability remains limited due to insufficient observational coverage in certain regions or variables, as well as the inadequate representation of the fine-scale physical processes responsible for local convective development. In response to these challenges, this study proposes a physically embedded neural network based on heterogeneous meteorological data, which utilizes satellite multispectral images and atmospheric temperature and humidity profile synergistically retrieved from space-based and ground-based infrared spectral observations, to forecast local Convective Initiation (CI) within a 6-hour lead time. The core innovation of this study lies in the development of a physically consistent model that explicitly embeds the Convective Available Potential Energy (CAPE) equation into the network architecture. By embedding physical information, the model enables the atmospheric thermodynamic feature extraction module to generate physically consistent feature tensors, thereby enhancing the representation of key convective processes. We trained the network using the pre-training and fine-tuning approach, then validated its effectiveness with reanalysis and actual observational data. The results demonstrate that incorporating the retrieved atmospheric profile data leads to a 40% improvement in the 6-hour average Critical Success Index (CSI), increasing from 0.44 to 0.62 relative to forecasts without atmospheric profile input. Furthermore, in validation experiments using reanalysis data and radar observations, the proposed atmospheric profile feature extraction module consistently improves the model's average forecast CSI by more than 29% compared to models utilizing purely data-driven profile extraction modules.